Reckoning Machines builds production systems for investment firms, with AI models inside the research process, and Monte Carlo decision models for corporate finance teams. The work stays traceable and auditable as it scales.
We build the system and run it alongside the team until it is handed over.
Investment firms
We have twice built a systematic equity platform and taken it through to live trading. The most recent one carries portfolio construction, optimization, order management and execution, with an LLM runner making about 11,000 calls per cycle inside the data pipeline. For a fundamental research team the same engineering puts AI models inside the research process with IRIS.
Corporate finance and FP&A
For a global payments company we built the profitability and unit-economics model for a new product and paired it with Monte Carlo scenario analysis. A large pharmaceutical company needed a way to decide which AI-enabled commercial investments to fund and which to test first. We built that decision framework on Monte Carlo and value-of-information analysis.
Point of view
As code gets cheaper, context and control matter more.
AI makes it easier to build and run forecasts, tests, and research agents. At scale, teams need to know which assumptions and evidence produced each result, what changed, and why a conclusion was accepted.
AI that proposes, a record of what it proposed, and a person who decides.
Built by Reckoning Machines
Two systems. One rule: audit the work.
IRIS is a harness for analysts: it preserves model context so AI can work more accurately with fewer tokens and analysts can trace the result. ARH is a harness for research agents: it tracks the search as alpha decays, keeping attempts, failures, and accepted findings visible. Audit gives both systems context and control.
ExcelClaude alongside
IRIS
IRIS turns a collection of financial models into a research system. Ask questions across your entire book. Understand what drives the numbers. Trace conclusions to their sources. Keep the intelligence as your coverage grows.
IRIS makes Claude for Excel smarter about your financial models, building toward cross-company investment intelligence.
Claude acts. IRIS observes, remembers, explains, and helps recover.
Analysts build and update Excel forecasts with Claude alongside them.
PMs see the whole book and ask how a question plays out across company models.
Follow the assumptions and calculations behind a forecast.
See what the numbers mean for the underlying business.
A Method is a forecasting skill with its evidence attached, and IRIS remembers every time you change it.
Return to the analyst's reasoning as the company and thesis change.
Review and approve proposed changes, with a record the team can inspect and rerun.
Semantic compression ties it together by giving Claude a compact account of the model, so it spends tokens on the question rather than on rediscovering the spreadsheet every time.
Analysts start from the IRIS Model Catalog rather than a blank sheet.
Supervised pilot with two investment firms. Access is arranged with us.
The Adaptive Research Harness holds research agents to account for what they try, and Dialog is the trading desk that runs on it.
Research agents now produce more ideas than a team can check, and the best-looking result of many attempts is not the same as one that will hold up.
ARH writes each attempt to the record before the next can start, with its question, its reserved budget, its result and its failures, so the count of attempts is a fact of the record rather than a recollection.
A separate evaluator checks each proposed conclusion, and a result enters the accepted record only under rules the owner has set.
Rejected work stays in the record, so the next researcher sees the failed variants and the decisions behind what was accepted.
Researchers and evaluators can be replaced while the acceptance rules stay in place, so a team can change its agents without losing what it has learned.
Dialog puts the harness to work on a paper desk. Competing research approaches develop hypotheses, an independent evaluator checks them, and the strategies that pass share one portfolio with orders, fills, cash and fees accounted for.
Budget makes the search visible and separate acceptance makes the result accountable, so when a strategy is challenged the history shows how much searching it took to find it and why it was accepted.
Open prototypes on GitHub. Synthetic data and paper execution.
Jed Gore, CFA, started as a software engineer on prime brokerage systems. He then spent nineteen years investing in financial-services stocks, five of them as portfolio manager of a $250 million portfolio at Diamondback Capital. Since 2017 he has built research and analytics platforms for investors. He has been the sole engineer on two systematic equity platforms taken to live trading, and ran the portfolio on one of them. He is a named inventor on a U.S. patent for extracting tables from financial filings.
Writing
Ideas argued in public before they ship:
Essays and worked examples behind our advisory practice and both systems.
